* feat(replication): purge delete markers by the target's own version id
When a delete marker is replicated, the target assigns it a version id. The
purge that follows derived one from the *source* uuid instead, which is only
correct when the target mirrors source version ids. A generic S3 target does
not: the derived id addresses a version that does not exist there, so the
purge is a no-op and the replica keeps a marker the source has already
removed. Same failure class as #4401.
Record the id the target reports and address it directly on purge.
Data path, all of it driven by the object's internal metadata rather than the
`ReplicationState` wire form, which encodes positionally and cannot carry a
map:
- `rustfs-utils`: the `replication-delete-marker-version-<arn>` key family,
plus `strip_internal_prefix_preserving_case` — ARNs are case-sensitive and
the existing `strip_internal_prefix` lowercases.
- `ReplicationState` gains the map and a `..._corrupt` flag, both
`#[serde(skip)]`; `ReplicatedTargetInfo` carries the per-target id.
- `persist_target_delete_marker_versions` is merge-only. A delete arriving
over internode RPC has an empty map, so treating it as authoritative would
let a remote disk erase an id the local disk still holds.
- `delete_object_version` copies the map into `fi.metadata` before dispatch,
so the durable carrier crosses the wire even though the field does not.
- The keys are folded into the quorum hash through their normalized form:
the dual internal prefixes carrying one mapping share an identity, while a
genuine disagreement between disks still shows up as a quorum difference.
- `corrupt` (the prefixes disagreed) fails closed: skip the purge and warn
rather than guess an id and risk destroying a live version on the target.
Ported from the rc.1 branch, which cannot merge as a whole: its MRF replay
rewrite collides with #5659/#5671/#5672/#5673 and regressed
`MRF_PENDING_CAP`. main's MRF machinery is kept; only this capability moves
across. It touches no MRF code.
Two things did not survive the port, deliberately. The branch's
`missing_is_complete` purge regression does not exist here — it came from its
own HEAD-precheck rewrite, and main's simpler path never had it. And the
branch's `MrfReplicateEntry` ordering fields are MRF-redesign scope, left
behind.
Verification: cargo fmt --all --check, git diff --check,
cargo check --workspace --all-targets, and the suites for the four touched
crates — 4070 tests, 2 pre-existing failures unrelated to this change
(`system_resolver_negative_result_reaches_the_dns_allowlist`,
`test_resolve_domain_preserves_system_resolver_error_provenance`; both are
the sandbox DNS interception, they fail on a clean checkout too).
* fix(replication): keep the layer guard happy
scripts/check_architecture_migration_rules.sh matches on text, so the doc
comments naming `rustfs_filemeta::` read as a cross-layer dependency even
though nothing imports it. Reword them; the guard passes.
* fix(replication): make the target-version cap deterministic
Two defects in this PR, both found in review.
The cap was applied while iterating a `HashMap`, so *which* 1000 entries
survived depended on iteration order. Two disks decoding the same oversized
metadata could keep different subsets, hash differently, and lose quorum —
instead of both reporting the same corruption. Collect first, then truncate
in `BTreeMap` order, which is total and identical everywhere.
And `persist_target_delete_marker_versions` discarded the `corrupt` flag from
the RPC carrier, committing a delete-marker update that looked clean while the
exact remote marker identity was unknown. It now declines to merge a corrupt
carrier. Because the helper only ever inserts, declining leaves the durable
keys already on the object untouched, which is strictly safer than writing a
mapping we cannot trust.
Residual, stated rather than papered over: corruption confined to the RPC
carrier is not persisted as a sentinel, so a later reader of an object that
carried no durable keys still sees "legacy, no mapping" rather than "corrupt".
Persisting that would need a wire-format addition; the consumer already fails
closed on any corruption it can observe.
New test: `target_delete_marker_versions_cap_is_deterministic_across_decodes`
decodes the same 1050-entry map twice and asserts both the corrupt flag and
the retained subset agree.
* fix(replication): preserve multipart source mtime (#5669)
* fix(kms): repair unopenable ciphertext and cover the Vault backends (#5668)
* Add black-box behavior tests for KMS resilience and serialization
* fix(kms): repair unopenable ciphertext across backends
Black-box testing of the KMS crate surfaced several defects that make
encrypted data permanently unreadable.
Symmetric envelopes. The Local and Vault Transit backends returned raw
cipher output from `encrypt` while `decrypt` parsed a JSON envelope, so
anything sealed through the master-key path could never be opened again.
Local also discarded the AES-GCM nonce. Both now emit the same envelope
`decrypt` consumes, matching the Static backend.
Deterministic AAD. The object layer derived AEAD additional data by
serializing a `HashMap` directly. Iteration order differs per instance,
so a context rebuilt from storage produced different AAD bytes than the
one used to seal and the object stopped opening. Ordering by key removes
that dependency, matching the Static backend's existing `context_aad`.
Objects written with the default single-key context are unaffected,
since a one-entry map has only one serialization.
Cipher in the header projection. `metadata_to_headers` recorded the SSE
mode (`AES256` / `aws:kms`), which cannot represent ChaCha20-Poly1305,
so a ChaCha-sealed object came back claiming `aws:kms` and was opened
with the wrong cipher. The cipher now travels in
`x-rustfs-encryption-algorithm` — the header the storage layer already
reads but nothing ever wrote. Objects without it fall back as before.
Also: the Static backend ignored `key_spec` and always issued 256-bit
data keys; Local `list_keys` hardcoded `truncated: false`, ignored
`marker`, and paginated over unordered `read_dir`, so a paginating
client silently saw a partial key list; and Local and Vault KV2 reported
`key_id: "unknown"` from `decrypt` despite the envelope naming the
master key.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(kms): cover both Vault backends and key rotation
The behavior suite ran only against Local and Static, and its own harness
documented the gap: the Vault backends had no business-capability
coverage at all. Setting `RUSTFS_KMS_VAULT_TOKEN` now adds Vault KV2 and
Vault Transit to every `for_each_backend` spec against a live server.
That lane is what surfaced the Transit envelope defect fixed in the
previous commit.
`rotate` and `versioning` are advertised only by the Vault backends, so
until now every capability-gated branch for them took the
`UnsupportedCapability` side and the working half was never asserted — a
rotation that dropped prior key versions would have gone green. The new
`behavior_rotation.rs` pins that half: material sealed before a rotation
still opens after it, repeated rotations accumulate versions rather than
overwriting a single spare, and the history survives a restart.
Two test defects fixed. `objects_round_trip_across_sizes_and_algorithms`
asserted a 1-byte object differs from its own ciphertext, which collides
once every 256 runs; the assertion now applies only where a collision is
not realistic, and small objects stay covered by the tag check and the
decrypt round-trip. `test_from_env_selects_token_file` depended on
`RUSTFS_KMS_VAULT_TOKEN` being absent from the caller's environment and
now clears it explicitly.
The snapshots directory was also removed from `.gitignore`: insta
snapshots are the assertions themselves, so leaving them untracked gives
CI nothing to compare against. Only `.snap.new` scratch files are
ignored now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(kms): adapt behavior suite to current key APIs
Rebasing onto main brought four API changes the suite predates.
`DeleteKeyRequest` gained `confirm_key_id`, and immediate deletion is now
gated on the server's `allow_immediate_deletion`. Scheduled deletions pass
`None`; the four specs that destroy a key outright echo the key id back
and opt the harness config in, which is what the gate asks of a real
caller.
`LocalBackupExportRequest` gained `sanitized_config`. These specs cover
the key-material path, so they seal no configuration and pass `None`.
`KmsCacheStats` became a named struct with real hit, miss, and eviction
counters. `cache_stats_returns_an_entry_count_and_no_hit_or_miss_data`
existed to pin the old placeholder behavior — that the second tuple
element was always zero — which main has since fixed, so it is now
`cache_stats_reports_hits_and_misses_separately` and asserts the counters
actually move.
Starting the service provisions the reserved probe key, so it shows up in
listings and backup bundles. Exact-set assertions filter it through a new
`without_probe_key` helper rather than naming it, keeping those specs
about the keys they seeded.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(kms): bind the AAD to the stored context bytes
Review caught that canonicalizing the AAD on decrypt breaks objects sealed
before canonicalization existed, and it was right. The AAD is the
*serialization* of the encryption context, and `x-rustfs-encryption-context`
stores that exact byte sequence: `encrypt_object` fed one `HashMap` to the
AEAD and then moved the same map into the metadata the header is written
from, so the stored string is byte-identical to the AAD the object was
sealed under. Those objects are therefore recoverable — but only while
nothing round-trips the value through a `HashMap` and re-serializes it.
Recomputing sorted AAD on decrypt would have turned a readable object into
a permanently unreadable one. The previous behavior was worse than the
first analysis credited: it did not merely fail intermittently, it made
the failure deterministic.
`EncryptionMetadata` now carries `context_aad`, the bytes the object was
actually sealed with. Encryption records what it fed the AEAD, the header
projection stores those bytes verbatim (and preserves a legacy ordering
across a re-projection rather than rewriting it into sorted form), and
`headers_to_metadata` carries the stored string through untouched. Both
decrypt paths, SSE-KMS and SSE-C, prefer it and fall back to canonical
serialization only when no stored serialization exists. Canonicalization
still applies to everything newly sealed, so the original ordering bug
cannot recur.
Two tests pin this: a legacy record whose sealed bytes are non-canonical
must survive a full header round trip unchanged, and a context header
rewritten to an equivalent-but-reordered serialization must fail
authentication rather than silently re-deriving a working AAD. Both were
mutation-checked against the reinstated bug on each side.
Also from review: the lifecycle churn test asserted only that every
request was accounted for, which holds whether the state gate exists or
not, so both branches are now pinned deterministically after the churn
(asserting `refused > 0` on the concurrent phase would only trade the hole
for a scheduling flake). And the Local and Vault KV2 envelopes compare
`encryption_context` without authenticating it — `DekCrypto` seals only
the plaintext — which is now documented at both sites; closing it needs a
versioned envelope, since existing ciphertext was sealed without AAD.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: ccccpj <ccccpj@outlook.com>
Co-authored-by: 唐小鸭 <tangtang1251@qq.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Preserve metadata replication operations in the durable MRF and route tagging, retention, and legal-hold updates through the existing full-object replication transport. Keep ACL propagation outside the contract because the current object model has no durable object ACL state.
Refs #1616
A same-name CopyObject marks the operation `metadata_only`, which lets the
store layer rewrite `xl.meta` in place and leave the data blocks untouched.
The handler independently strips the source encryption metadata and calls
`sse_encryption`, which mints a *fresh* DEK. On an unversioned bucket both
happen at once, so the object ends up with a new DEK sitting beside ciphertext
sealed under the old one, and can never be decrypted again.
The mirror case is silent: an encrypted source copied without any destination
SSE keeps its ciphertext while losing the key metadata, so GET returns raw
ciphertext as if it were plaintext, with HTTP 200 and no error anywhere.
Keep `metadata_only` off whenever either side of the copy is encrypted, so the
store layer performs a full read/write rewrite through `put_object`. This is
the same resolution the versioned historical-restore path already uses for
this risk (issue #4238), and it matches MinIO's
`isSourceEncrypted || isTargetEncrypted -> metadataOnly = false` guard in
CopyObjectHandler.
The target half of the predicate deliberately tests `effective_sse` rather
than the request headers MinIO inspects: `effective_sse` also resolves the
bucket default-encryption rule, and `sse_encryption` mints a DEK from that
resolved value. A header-only check would miss a same-key copy performed under
a bucket default rule. The source half reuses `ObjectInfo::is_encrypted` so a
future encryption flavour is covered here as soon as it is recognised there.
Versioned buckets were already safe: that path falls through to `put_object`
regardless of `metadata_only`. RestoreObject also sets `metadata_only` but
only appends restore keys and never re-derives a DEK, so it is unaffected.
Add RAII guards for replication runtime backlog tickets so active worker and queue counters unwind on every terminal path.
Expose node-local MRF pending, dropped, missed, and flush-failure metrics through the bucket replication Prometheus collector while keeping the existing current backlog and durable MRF gauges additive.
Update durable MRF summary maintenance to aggregate incrementally during the persister loop, avoiding repeated full-entry scans on each successful flush.
Co-authored-by: heihutu <heihutu@gmail.com>
Co-authored-by: zhi22915 <qiuzgang@gmail.com>
A streaming GET holds snapshot leases on the object's data directories,
and DeleteObjects defers their physical cleanup until the leases are
released. A DeleteBucket issued inside that window passes the xl.meta
emptiness check but fails closed in the non-force delete_volume tree
removal on the leftover part files, returning BucketNotEmpty for a
logically empty bucket.
This is what intermittently failed the s3tests
test_encryption_sse_c_multipart_bad_download teardown in CI: the test
never reads its 30MiB GET body, so the server-side stream (and its
leases) stays alive until the connection drops, racing the teardown's
DeleteObjects + DeleteBucket sequence. With the body held open the
failure reproduces 5/5 locally; after this change it passes 20/20, and
the real s3-tests case passes 20 consecutive runs.
delete_volume now executes the registry-tracked pending deferred
deletions for the volume before removing the directory tree. Only data
dirs whose logical delete already committed are touched; unknown files
still fail closed with VolumeNotEmpty.
Add opt-in hotpath feature surfaces to every workspace crate and wire the root rustfs feature passthrough for function, allocation, and CPU profiling.
Add a focused set of function-level measurements for scanner, heal, lock, target replay, IAM, KMS, Keystone, trusted proxy, and capacity paths without adding request-scoped primitive wrappers.
Co-authored-by: heihutu <heihutu@gmail.com>
* refactor(sse): decouple encryption from ecstore
* feat(kms): enhance KMS service manager with runtime state and persistence support
* feat(kms): add local key export functionality for SSE-S3 migration tests
* fix(kms): keep local key export narrowly scoped
* fix(sse): validate copy source customer algorithm
---------
Co-authored-by: Zhengchao An <anzhengchao@gmail.com>
rename_all_missing_source_still_warns and rename_all_real_failure_still_warns
assert that `warn_reliable_rename_failure` emitted its WARN, but that is a
single production callsite shared with tests that call rename_all *without*
installing a subscriber — rename_all_missing_source_returns_file_not_found,
two tests above, is one of them.
tracing caches each callsite's Interest process-globally and the first thread
to reach a callsite fixes that value; while at most one dispatcher is
registered, tracing-core derives it from the registering thread's own
subscriber, and registration is once-only. When the subscriber-less sibling
wins, the callsite is cached as Interest::never() and the WARN never fires,
so the assertion sees empty output:
ordinary missing-source failures must keep the WARN, got:
Reproduced at 3/25 with `disk::os::tests::rename_all_missing_source` (both
tests), against 0/20 for the victim alone. Fixed by pinning callsite interest
inside warn_capture(), so every current and future user of that helper is
covered rather than just the two tests that happen to fail today.
pin_callsite_interest_for_test() moves from cluster::rpc::background_monitor
to a new crate-level test_tracing module: it is domain-neutral and now has
consumers in two unrelated subsystems, and disk::os should not have to reach
into a cluster::rpc test helper.
Verified: repro filter 0/30 (was 3/25); disk::os:: 0/12; cluster::rpc:: 0/12
and its poisoner pair 0/15, confirming the moved helper still holds.
Follow-up to #5438. Closes the last item in #5439.
Only replication-target writes took the bucket transaction lock. Every other
config write (policy, tagging, lifecycle, versioning, ...) went straight to
the process-local metadata-system guard, which serializes nothing across
nodes.
Each config write is a read-modify-write of one whole BucketMetadata blob:
load the blob, replace one field, save the blob back. The namespace locks
inside read_config/save_config are taken and released separately, so they do
not span that cycle. Two nodes updating different config files of the same
bucket therefore both load the same blob, each set their own field, and the
later save drops the other's -- with both clients already told 2xx. This is
not last-writer-wins on one document; an orthogonal config silently vanishes.
Route update(), delete() and update_config_with() through
acquire_config_write_guards(), which takes the cluster-wide transaction lock
first and the metadata-system write guard second. That order is load-bearing:
taking the process-local guard first would park every local reader and writer
of every bucket behind a lock whose holder may be another node, turning
remote contention into a local stall.
The lock is per bucket rather than per config file, since a per-file key
would let exactly the offending pair run concurrently. Rename the helper to
acquire_bucket_metadata_transaction_lock to match, but deliberately keep the
lock resource string as "bucket-targets/{bucket}/transaction.lock": the key
is what nodes agree on, so renaming it would leave a mixed-version cluster
with two disjoint keys and stop old and new nodes from excluding each other
on the very writes that are serialized today.
update_config_with() already narrowed its staleness window to a single load
and save, but its exclusion was explicitly process-local; it is now
cluster-wide, so its doc comment no longer disclaims cross-node races.
Also make update_and_parse load through self.api instead of the ambient store
handle, so the read and the write of one read-modify-write cannot resolve to
different instances.
The new tests drive two BucketMetadataSys instances over one ECStore -- the
in-process stand-in for two nodes, since they share no RwLock and can only be
serialized by the namespace lock. Verified the lost-update test has teeth by
removing the lock and confirming it fails on round 0, with the tagging config
clobbered to empty by the concurrent policy write.
peer_rest_recovery_probe_logs_keep_request_id_span_context failed ~10% of
`cargo test -p rustfs-ecstore --lib -- cluster::rpc::` runs with
left: "request-span", right: "recovery-monitor". The recovery-monitor
info_span! was evaluating to Span::none(), so the probe's log line landed
under the caller's span.
tracing caches each callsite's Interest in process-global state, and the
first thread to reach a callsite fixes that value. While at most one
dispatcher is registered, tracing-core takes a fast path that derives the
interest from the registering thread's own subscriber, and registration is
once-only (CAS). Under libtest a sibling test reaches recovery_monitor_span
via mark_offline_and_spawn_recovery from a thread with no subscriber, so
the interest is derived from NoSubscriber and cached as Interest::never()
for the whole process.
Add pin_callsite_interest_for_test(): registering a second, inert
dispatcher rebuilds every registered callsite's interest against the live
dispatcher set (repairing a poisoned value) and keeps tracing-core off the
single-dispatcher fast path (preventing new ones). This also covers the
production marked_suspect / recovery_monitor_started event callsites that
remote_disk_network_error_starts_recovery_monitor_with_request_context
asserts on.
rename_data_response_accepts_legacy_json_without_decode_error is a
separate root cause: it snapshots the process-global internode metrics and
asserts the decode-error counter did not move, which siblings that record
decode errors (or reset the counters) invalidate. Put the 11 tests that
observe those counters in one #[serial(internode_metrics)] group.
Both races are impossible under nextest, which runs each test in its own
process, so neither test belongs in the ecstore-serial-flaky test-group
(that serializes across process boundaries) nor in the ci-profile
quarantine (they never redden CI).
Verified: cluster::rpc:: subset 0/30 failures under libtest (was 3/30);
target test paired with its poisoner 0/30 (was 4/20); 5/5 clean under
nextest at 179/179.
Zhengchao An